swap to Qwen2.5-VL-3B (sweet spot: better vision than SmolVLM, fits ZeroGPU)

app.py

@@ -10,24 +10,23 @@ from PIL import Image
 import json
 import re
 import torch
-from transformers import AutoProcessor, AutoModelForVision2Seq
+from transformers import AutoProcessor, AutoModelForImageTextToText
 
 # ZeroGPU integration — no-op outside HF Spaces
 try:
     import spaces
 
     def gpu(fn):
-        return spaces.GPU(duration=60)(fn)
+        return spaces.GPU(duration=90)(fn)
 except ImportError:
     def gpu(fn):
         return fn
 
 
-# SmolVLM-Instruct (2.25B). Bigger than the 500M we tested earlier —
-# much better at fine-grained vision tasks like spotting queens.
-# Qwen2.5-VL-7B kept crashing the ZeroGPU container at the OCI runtime
-# layer; reverting to the SmolVLM family which is known to work here.
-MODEL_ID = "HuggingFaceTB/SmolVLM-Instruct"
+# Qwen2.5-VL-3B. State-of-the-art vision-language model at a size that
+# comfortably fits on ZeroGPU. SmolVLM-2.25B hallucinated on bee anatomy;
+# Qwen-7B crashed the container; this is the sweet spot.
+MODEL_ID = "Qwen/Qwen2.5-VL-3B-Instruct"
 
 _model = None
 _processor = None
@@ -39,7 +38,10 @@ def get_model():
     if _model is None:
         print(f"Loading {MODEL_ID} ...")
         _processor = AutoProcessor.from_pretrained(MODEL_ID)
-        _model = AutoModelForVision2Seq.from_pretrained(MODEL_ID)
+        _model = AutoModelForImageTextToText.from_pretrained(
+            MODEL_ID,
+            torch_dtype=torch.float16,
+        )
         _model.eval()
         print("Model loaded.")
     return _model, _processor
@@ -105,7 +107,7 @@ def build_narrative(r: dict, raw: str) -> str:
 **Notes:** {r['notes']}
 
 ---
-*Powered by SmolVLM-Instruct (2.25B) on ZeroGPU. Fully local, no cloud APIs.*
+*Powered by Qwen2.5-VL-3B on ZeroGPU. Fully local, no cloud APIs.*
 
 <details><summary>Raw model output</summary>
 
@@ -131,17 +133,20 @@ def analyze_frame(image: Image.Image, hive_name: str):
         {
             "role": "user",
             "content": [
-                {"type": "image"},
+                {"type": "image", "image": image},
                 {"type": "text", "text": INSPECTION_PROMPT},
             ],
         }
     ]
 
     try:
-        prompt = processor.apply_chat_template(
-            messages, add_generation_prompt=True
+        inputs = processor.apply_chat_template(
+            messages,
+            tokenize=True,
+            add_generation_prompt=True,
+            return_dict=True,
+            return_tensors="pt",
         )
-        inputs = processor(text=prompt, images=[image], return_tensors="pt")
         inputs = inputs.to(device)
 
         with torch.no_grad():
@@ -236,7 +241,7 @@ with gr.Blocks(title="Apiarist - Hive Frame Inspector") as app:
 - 🎯 Vision-language model fine-tuned for honeybees
 - 📓 Built in 10 days for the [Build Small Hackathon](https://huggingface.co/build-small-hackathon)
 
-**Stack**: SmolVLM-Instruct (2.25B) on ZeroGPU, served via Gradio.
+**Stack**: Qwen2.5-VL-3B on ZeroGPU, served via Gradio.
 """
             )